ARM + TSMC leading Intel In SOC Process Technology, says East.

ARM + TSMC is leading Intel in process technology for highly integrated mobile SOCs, ARM’s CEO Warren East tells me.

Asked if it was the aspiration of ARM’s physical IP division, working with TSMC, to catch up with Intel in process, East replies: “Implicit in your question is that we’re behind Intel. Intel is doing 22nm for Ivy Bridge but, when it comes to mobile SOCs, ARM + TSMC are doing 28nm planar CMOS for SOC, while Intel is using a 32nm metal gate process for SOC. We want to have highly integrated SOC devices. These are more difficult to put together than microprocessors.”

ARM is working with TSMC to develop optimised processes for ARM-based SOCs. “Our mutual customers want to create products competitive with the Intel world,” says East. “with ARM + TSMC customers have an optimised solution. We’re optimising our physical IP working up to system design, and down to the transistors with POPs (Processor Optimisation Packs) attuned to semiconductor process technology.”

Asked what he expected the performance of ARM-based Windows 8 computers would be like, East replies: “There are hints in the Microsoft Surface. There are two versions, one ARM, one x86. The ARM version weighed two thirds less, it was two thirds of the thickness and it didn’t have cooling vents. It suggests to me that it’s a lower power design.”

Next year, Intel is intending to put its Atom processors onto its 22nm finfet process which may improve Atom’s power efficiency. It’s got some way to go. “An 800MHz ARM delivers the same performance as a 1.6GHz Atom,” says East.

Asked if ARM’s 64-bit architecture is for servers, East says: “Yes but we’re also finding interest in the networking community, in mobile phone apps processors and from mobile computing companies.”

TSMC is moving to finfets at 20nm which will change the physical IP requirements, which is where ARM is working particularly closely with TSMC, but won’t change the processor design.

Asked how finfets would affect the performance of ARM cores, East says: “The performance has little to do with us. It depends on what the foundry is expecting.”

TSMC’s expectations for planar 20nm over 28nm are a 1.9x increase in density, 25% less power requirement and 15-20% more performance.

Presumably TSMC wouldn’t be introducing finfets at 20nm unless its performance characteristics are better than that.



  1. Well, raisinhead, it’s a bit about process, and it’s a bit, as you say, about the different business models but it’s also about the difference between making microprocessors on the one hand and stitching together complex SOCs on the other hand. The difficulties involved in the latter are often under-estimated.

  2. Although the article is really a case of show-boating, Intel vs ARM is about so much more than process technology. It is is Intel vs a whole bunch or people ARM/Broadcom/Qualcomm/Marvell/etc. It is the monolithic vs the diverse. It is a closed model vs an open model.
    An analogy would be Android vs Windows Phone.
    ARM IP model allows the customers greater share the revenue in the value chain and allows them greater opportunity of differentiation – much more important to un-Intel guys than squeezing last drops of performance out of a process.
    ARM is much smaller than Intel as a result of course. But look at the combined revenues of the names mentioned. A powerful coalition of the unIntel-igentsia (I know – ugly syllogism).

  3. Azzouz Nezar writes: A while back I suspected Intel will fail with their Atom for SoC considering my experience with Intel. Intel finfet and even planar are tuned exactly to their x86 and cannot accomodate complex devices needed for SoC. They do not care whether the device model are wrong! They do not have to deal with analog or any embedded memory (Flash or other), etc.

  4. Well the data I have is that both Intel and TSMC produce approximately 260,000 300mm wafers per month at 45nm or better.
    However Intel make 210,000 at 32nm or better whilst TSMC only make 130,000 at 32nm or better. Of course both companies are raising these numbers all the time so they are just a snapshot from earlier this year.
    But if we only count the true state of the art then Intel produce > 100,000 wpm at 22nm and TSMC produce …. some test wafers.
    So your statement “in terms of actual produced wafers they don’t come close to tsmc among several others” is incorrect.
    You can also add to this the fact that Intel uses two processes – 22nm Trigate and 32nm HKMG, whilst TSMC seem to have more processes than they know what to do with. Fortunately at 20nm TSMC have seen sense and reduced to the correct number – 1 !! This will improve their yields significantly.
    Of course TSMC produce huge numbers of larger geometries but this whole thread has been considering the state of the art. And in any case Intel still has capacity of over 200,000 wpm at these larger nodes which it is slowly closing down.
    And whilst mobile SoCs use less layers, products for AMD, nVidia, Altera and Xilinx are pretty layer intensive as well.

  5. “simply because Intel makes more wafers than anyone else”
    Excuse me?
    Intel may make more revenue than most due to the price per chip, but in terms of actual produced wafers they don’t come close to tsmc among several others.
    Look at their litho capex, then remember Intel uses more layers than most other semi manufacturers (for processors anyway) and they have by far the lowest utilization of any fab, then count your wafers again.

  6. Ah, sorry I got the wrong end of the stick, Cassandra. Well as the Americans say: ‘What goes around comes around’ and if you recall what Intel and the microprocessor did to the mainframe and mini-computer industry, it’s not beyond the bounds of possibility that multiple sources of computer CPUs could do to Intel what Intel did to the mainframe industry. Intel’s size totally depends on maintaining its monopoly position in computer CPUs. Without the monopoly it could become a very poor thing.

  7. Hi David, I was not clear. I was not thinking of Nokia just as a phone maker, but as an example of an undisputed giant in the high-tech industry, experiencing a sudden change of fortunes, actually a spiral down to … hell? 😉

  8. If it is, Cassandra, it’s a well disguised plan. So far Intel’s only got its mobile SOCs into handset makers with funny names like Lolly and Zolo or somesuch which no one’s ever heard of. Mind you, with enough Intel MDF money, maybe Zoilly and and Lolo will be the next Nokia and RIM.

  9. David, is Intel waiting in line to be the next Nokia or the next RIM ?

  10. We’ll see. As a humble journo I don’t have to forecast anything, thank goodness, I just wait and wonder.

  11. Well ARM and Qualcomm had better not rest on their laurels or it will be an Intel monopoly 🙂
    Whilst Qualcomm is frantically looking for the best way forward the messages out of ARM, and sometimes TSMC as well, are far too complacent for my liking.
    And the biggest mistake was deciding to produce the 20nm planar process as this will delay TSMC getting the finfet process needed to compete with Intel to market.

  12. Oh well we were talking at cross-purposes. I was stuck in the here and now. As for the future: never forget the semiconductor industry’s Golden Rule: ‘When Intel steps outside its monopoly it always screws up’.

  13. Not at all. I never disputed that the current Medfield product is on 32nm and accept this lags just behind current 28nm TSMC fabbed products. However your article was on future products – you mention TSMC 20nm and Finfets, ARMv8 and other things still a long way off whereas Intel’s 22nm SoC is aimed for full release next April. The fact this uses existing fab processes takes the risk out of it – Intel’s tick-tock approach – and so it will be very surprising if this slips.
    Intel’s rumoured yield figures – they never report them – are dominated by huge PC processor cores. The much smaller Atom processors will have yields better than industry norms simply because Intel makes more wafers than anyone else and so can optimise yield for a single process. AMD has tried to compete at this level and always failed on cost.
    There will however be a huge step forward for TSMC at 20nm when they stop offering lots of different processes as they too will gain the experience of optimising a single process using data from all of their customer’s wafers.
    And as I said, cost will be whatever Intel decide to charge for it. They can always afford to undercut anyone else should they choose to do as we have always argued having in-house production will always be cheaper than using foundries and you have reported and generally agreed with this thesis.

  14. Well if Intel’s 22nm mobile SOC is not even designed yet, then Warren is absolutely correct in saying Intel’s current commercial mobile SOCs are on 32nm. Moreover, if it’s not even designed yet, the ‘huge’ performance improvement is problematic. Furthermore Intel’s yields have always been reported to be lower than anyone else’s because Intel cares less about production cost than anyone else because it enjoys monopoly pricing. While, in the mobile SOC industry, which is not a monopoly industry, Intel won’t get away with charging whatever it wants for it.

  15. You know full well it’s about three quarter way through design, which is a lot further on than any designs using TSMC’s 20nm non-finfet process, let alone the promised finfet process.
    Performance will be a huge improvement over Medfield and Cortex A9 designs, and Intel’s yield will be as good as ever. Cost will be whatever Intel decide to charge for it.

  16. And their non-appearance on the market is due to their inferior performance characteristics? Lousy yield? Uncompetitive cost?

  17. They are.

  18. I knew, when I wrote it, that I’d get a snotty remark from you, Mike, so: if your Intel pals are so good at mobile SOCs why haven’t they made one on their 22nm finfet process?

  19. I see you’re entering for the Booker Prize this year.

Leave a Reply

Your email address will not be published. Required fields are marked *